Research Center for Digital Technologies in Dentistry and CAD/CAM, Department of Dentistry, Faculty of Medicine and Dentistry, Danube Private University, Steiner Landstraße 124, 3500, Krems, Austria.
Department of Paediatric Dentistry, Medical University of Vienna - University Clinic of Dentistry, Sensengasse 2a, 1090, Vienna, Austria.
Clin Oral Investig. 2022 Dec;26(12):6917-6923. doi: 10.1007/s00784-022-04646-z. Epub 2022 Sep 6.
Molar incisor hypomineralization (MIH) is a difficult-to-diagnose developmental disorder of the teeth, mainly in children and adolescents. Due to the young age of the patients, problems typically occur with the diagnosis of MIH. The aim of the present technical note was to investigate whether a successful application of a neural network for diagnosis of MIH and other different pathologies in dentistry is still feasible.
For this study, clinical pictures of four different pathologies were collected (n = 462). These pictures were categorized in caries (n = 118), MIH (n = 115), amelogenesis imperfecta (n = 112) and dental fluorosis (n = 117). The pictures were anonymized and a specialized dentist taking into account all clinical data did the diagnosis. Then, well-investigated picture classifier neural networks were selected. All of these were convolutional neural networks (ResNet34, ResNet50, AlexNet, VGG16 and DenseNet121). The neural networks were pre-trained and transfer learning was performed on the given datasets.
For the vgg16 network, the precision is the lowest with 83.98% as for the dense121 it shows the highest values with 92.86%. Comparing the different pathologies between the investigated neural networks, there is no trend detectable.
In the long term, an implementation of artificial intelligence for the detection of specific dental pathologies is conceivable and sensible.
Finally, this application can be integrated in the area of training and teaching in order to teach dental students as well as general practitioners for MIH and similar dental pathologies.
摩尔牙本质发育不全（MIH）是一种难以诊断的牙齿发育障碍，主要发生在儿童和青少年中。由于患者年龄较小，通常会在 MIH 的诊断中出现问题。本技术说明的目的是研究用于诊断 MIH 和其他不同牙科学病症的神经网络的成功应用是否仍然可行。
为此研究，收集了四种不同病症的临床图片（n=462）。这些图片分为龋病（n=118）、MIH（n=115）、釉质发育不全（n=112）和氟牙症（n=117）。图片进行了匿名化处理，由一位考虑所有临床数据的专业牙医进行诊断。然后，选择了经过充分研究的图片分类器神经网络。所有这些都是卷积神经网络（ResNet34、ResNet50、AlexNet、VGG16 和 DenseNet121）。对神经网络进行了预训练，并在给定的数据集上进行了迁移学习。
对于 vgg16 网络，精度最低，为 83.98%，而对于 dense121，精度最高，为 92.86%。比较不同病症之间的不同神经网络，没有可检测到的趋势。
从长远来看，人工智能在检测特定牙科学病症方面的应用是可以想象的，也是合理的。
最后，该应用可以集成到培训和教学领域，以便教授牙科学生以及一般从业者有关 MIH 和类似牙科学病症的知识。
